Executive Summary
CytoAtlas is a comprehensive computational resource that maps cytokine and secreted protein signaling activity across 240 million human cells from six independent datasets spanning healthy donors, inflammatory diseases, cancers, drug perturbations, and spatial transcriptomics. The system uses linear ridge regression against experimentally derived signature matrices to infer activity — producing fully interpretable, conditional z-scores rather than black-box predictions.
Key results:
- 1,293 signatures (44 CytoSig + 178 LinCytoSig + 1,249 SecAct) validated across 8 independent atlases
- Spearman correlations reach ρ=0.6–0.9 for well-characterized cytokines (IL1B, TNFA, VEGFA, TGFB family)
- Cross-atlas consistency demonstrates signatures generalize across CIMA, Inflammation Atlas, scAtlas, GTEx, and TCGA
- LinCytoSig improves prediction for select immune cell types (Basophil, NK, DC: +0.18–0.21 Δρ)
- SecAct achieves the highest correlations in bulk & organ-level analyses (median ρ=0.40 in GTEx/TCGA)
Table of Contents
1. System Architecture and Design Rationale
1.1 Why This Architecture?
CytoAtlas was designed around three principles that distinguish it from typical bioinformatics databases:
Principle 1: Linear interpretability over complex models.
Ridge regression (L2-regularized linear regression) was chosen deliberately over methods like autoencoders, graph neural networks, or foundation models. The resulting activity z-scores are conditional on the specific genes in the signature matrix, meaning every prediction can be traced to a weighted combination of known gene responses.
Principle 2: Multi-level validation at every aggregation.
CytoAtlas validates at five levels: donor-level pseudobulk, donor × cell-type pseudobulk, single-cell, bulk RNA-seq (GTEx/TCGA), and bootstrap resampled with confidence intervals.
Principle 3: Reproducibility through separation of concerns.
| Component | Technology | Purpose |
|---|---|---|
| Pipeline | Python + CuPy (GPU) | Activity inference, 10–34x speedup |
| Storage | DuckDB (3 databases, 68 tables) | Columnar analytics, no server needed |
| API | FastAPI (262 endpoints) | RESTful data access, caching, auth |
| Frontend | React 19 + TypeScript | Interactive exploration (12 pages) |
1.2 Processing Scale
| Dataset | Cells | Samples | Time | GPU |
|---|---|---|---|---|
| CIMA | 6.5M | 421 donors | ~2h | A100 |
| Inflammation Atlas | 6.3M | 1,047 samples | ~2h | A100 |
| scAtlas | 6.4M | 781 donors | ~2h | A100 |
| parse_10M | 9.7M | 1,092 conditions | ~3h | A100 |
| Tahoe-100M | 100.6M | 14 plates | ~12h | A100 |
| SpatialCorpus-110M | ~110M | 251 datasets | ~12h | A100 |
2. Dataset Catalog
2.1 Datasets and Scale
| # | Dataset | Cells | Donors/Samples | Cell Types | Reference |
|---|---|---|---|---|---|
| 1 | CIMA | 6,484,974 | 421 donors | 27 L2 / 100+ L3 | J. Yin et al., Science, 2026 |
| 2 | Inflammation Atlas | 6,340,934 | 1,047 samples | 66+ | Jimenez-Gracia et al., Nature Medicine, 2026 |
| 3 | scAtlas | 6,440,926 | 781 donors | 100+ | Q. Shi et al., Nature, 2025 |
| 4 | parse_10M | 9,697,974 | 12 donors × 91 cytokines | 18 PBMC types | Oesinghaus et al., bioRxiv, 2026 |
| 5 | Tahoe-100M | ~100,600,000 | 50 cell lines × 95 drugs | 50 cell lines | Zhang et al., bioRxiv, 2026 |
| 6 | SpatialCorpus-110M | ~110,000,000 | 251 spatial datasets | Variable | Tejada-Lapuerta et al., Nature Methods, 2025 |
2.2 Disease and Condition Categories
Inflammation Atlas (20 diseases): RA, SLE, Sjogren's, PSA, Crohn's, UC, COVID-19, Sepsis, HIV, HBV, BRCA, CRC, HNSCC, NPC, COPD, Cirrhosis, MS, Asthma, Atopic Dermatitis
scAtlas: Normal (35+ organs) + Cancer (15+ types: LUAD, CRC, BRCA, LIHC, PAAD, KIRC, OV, SKCM, GBM, etc.)
parse_10M: 90 cytokines × 12 donors — independent in vitro perturbation dataset for comparison. A considerable portion of cytokines (~58%) are produced in E. coli, with the remainder from insect (Sf21, 12%) and mammalian (CHO, NS0, HEK293, ~30%) expression systems. Because exogenous perturbagens may induce effects differing from endogenously produced cytokines, parse_10M serves as an independent comparison rather than strict ground truth. CytoSig/SecAct has a potential advantage in this regard, as it infers relationships directly from physiologically relevant samples.
Tahoe-100M: 95 drugs across 50 cancer cell lines
SpatialCorpus: Visium, Xenium, MERFISH, MERSCOPE, CosMx, ISS, Slide-seq — 30+ tissue types
2.3 Signature Matrices
| Matrix | Targets | Construction | Reference |
|---|---|---|---|
| CytoSig | 44 cytokines | Median log2FC across all experimental bulk RNA-seq | Jiang et al., Nature Methods, 2021 |
| LinCytoSig | 178 (45 cell types × 1–13 cytokines) | Cell-type-stratified median from CytoSig database (methodology) | This work |
| SecAct | 1,249 secreted proteins | Median global Moran's I across 1,000 Visium datasets | Ru et al., Nature Methods, 2026 (in press) |
3. Scientific Value Proposition
3.1 What Makes CytoAtlas Different from Deep Learning Approaches?
Most single-cell analysis tools use complex models (VAEs, GNNs, transformers) that produce aggregated, non-linear representations difficult to interpret biologically. CytoAtlas takes the opposite approach:
| Property | CytoAtlas (Ridge Regression) | Typical DL Approach |
|---|---|---|
| Model | Linear (z = Xβ + ε) | Non-linear (multi-layer NN) |
| Interpretability | Every gene's contribution is a coefficient | Feature importance approximated post-hoc |
| Conditionality | Activity conditional on specific gene set | Latent space mixes all features |
| Confidence | Permutation-based z-scores with CI | Often point estimates only |
| Generalization | Tested across 8 independent cohorts | Often held-out splits of same cohort |
| Bias | Transparent — limited by signature matrix genes | Hidden in architecture and training data |
The key insight: CytoAtlas is not trying to replace DL-based tools. It provides an orthogonal, complementary signal that a human scientist can directly inspect. When CytoAtlas says "IFNG activity is elevated in CD8+ T cells from RA patients," you can verify this by checking the IFNG signature genes in those cells.
3.2 What Scientific Questions Does CytoAtlas Answer?
- Which cytokines are active in which cell types across diseases?
- Are cytokine activities consistent across independent cohorts?
- Does cell-type-specific biology matter for cytokine inference?
- Which secreted proteins beyond cytokines show validated activity?
- How do drugs alter cytokine activity in cancer cells?
- What is the spatial organization of cytokine signaling?
- Can we predict treatment response from cytokine activity?
3.3 Validation Philosophy
CytoAtlas validates against a simple but powerful principle: if CytoSig predicts high IFNG activity for a sample, that sample should have high IFNG gene expression. This expression-activity correlation is computed via Spearman rank correlation across donors/samples.
This is a conservative validation — it only captures signatures where the target gene itself is expressed. Signatures that act through downstream effectors would not be captured, meaning our validation underestimates true accuracy.
4. Validation Results
4.1 Overall Performance Summary
How “N Targets” is determined: A target is included in the validation for a given atlas only if (1) the target’s signature genes overlap sufficiently with the atlas gene expression matrix, and (2) the target gene itself is expressed in enough samples to compute a meaningful Spearman correlation. Targets whose gene is absent or not detected in a dataset are excluded.
Donor-only atlases (CIMA, Inflammation, GTEx, TCGA): N = number of unique targets with valid correlations. CytoSig defines 43 cytokines and SecAct defines 1,170 secreted proteins. The Inflammation Atlas (main/validation cohorts) retains only 33 of 43 CytoSig targets and 805 of 1,170 SecAct targets because 10 cytokine genes (BDNF, BMP4, CXCL12, GCSF, IFN1, IL13, IL17A, IL36, IL4, WNT3A) are not sufficiently expressed in these blood/PBMC samples. CIMA, GTEx, and similar multi-organ datasets retain nearly all targets (≥97%).
Donor-organ atlases (scAtlas Normal, scAtlas Cancer): N = target × organ pairs, because validation is stratified by organ/tissue context. For scAtlas Normal, each target is validated independently across 25 organs (Bladder, Blood, Breast, Colon, Heart, Kidney, Liver, Lung, etc.), yielding up to 43 × 25 = 1,075 CytoSig entries (actual: 1,013 after filtering) and 1,140 × 25 = 28,500 SecAct entries (actual: 27,154). For scAtlas Cancer, validation spans 7 tissue contexts (Tumor, Adjacent, Blood, Metastasis, Pleural Fluids, Pre-Lesion, All), yielding 43 × 7 = 301 CytoSig entries (actual: 295) and 1,140 × 7 = 7,980 SecAct entries (actual: 7,809). Some target-organ pairs are excluded when the target gene lacks sufficient expression in that organ.
Note on scAtlas duplicate entries: At finer aggregation levels (e.g., donor_organ_celltype1 vs donor_organ_celltype2), the same target can appear multiple times with different correlation values. This is expected — finer cell-type annotation changes the composition of each pseudobulk sample, yielding different expression-activity relationships. The summary table above uses the donor_organ level for scAtlas.
4.2 Correlation Distributions
Why does SecAct appear to underperform CytoSig in the Inflammation Atlas?
This is a composition effect, not a genuine performance gap. CytoSig tests only 43 curated, high-signal cytokines, while SecAct tests 1,249 secreted proteins — including many tissue-expressed targets (collagens, metalloproteinases, apolipoproteins, complement factors) with minimal expression variation in blood/PBMC samples. On the 22 matched targets shared between both methods, SecAct consistently outperforms CytoSig across all atlases (e.g., median ρ = 0.51 vs 0.32 in Inflammation Main).
The Inflammation Atlas is largely blood-derived, so many SecAct targets that perform well in multi-organ contexts (scAtlas, GTEx, TCGA) contribute near-zero or negative correlations here. In fact, 99 SecAct targets are negative only in inflammation but positive in all other atlases, reflecting tissue-specific expression limitations rather than inference failure. The “SecAct (CytoSig-matched)” boxplot above demonstrates the fair comparison on equal footing.
4.3 Best and Worst Correlated Targets
Consistently well-correlated targets (ρ > 0.3 across multiple atlases):
- IL1B (ρ = 0.67 CIMA, 0.68 Inflammation) — canonical inflammatory cytokine
- TNFA (ρ = 0.63 CIMA, 0.60 Inflammation) — master inflammatory regulator
- VEGFA (ρ = 0.79 Inflammation, 0.92 scAtlas) — angiogenesis factor
- TGFB1/2/3 (ρ = 0.35–0.55 across atlases)
- BMP2/4 (ρ = 0.26–0.92 depending on atlas)
Consistently poorly correlated targets (ρ < 0 in multiple atlases):
- CD40L (ρ = −0.48 CIMA, −0.56 Inflammation) — membrane-bound, not secreted
- TRAIL (ρ = −0.46 CIMA, −0.55 Inflammation) — apoptosis inducer
- LTA (ρ = −0.33 CIMA), HGF (ρ = −0.25 CIMA)
Gene mapping verified: All four targets are correctly mapped (CD40L→CD40LG, TRAIL→TNFSF10, LTA→LTA, HGF→HGF). No gene ID confusion exists. The poor correlations reflect specific molecular mechanisms:
| Target | Gene | Dominant Mechanism | Contributing Factors |
|---|---|---|---|
| CD40L | CD40LG | Platelet-derived sCD40L invisible to scRNA-seq (~95% of circulating CD40L); ADAM10-mediated membrane shedding | Unstable mRNA (3′-UTR destabilizing element); transient expression kinetics (peak 6–8h post-activation); paracrine disconnect (T cell → B cell/DC) |
| TRAIL | TNFSF10 | Three decoy receptors (DcR1/TNFRSF10C, DcR2/TNFRSF10D, OPG/TNFRSF11B) competitively sequester ligand without signaling | Non-functional splice variants (TRAIL-beta, TRAIL-gamma lack exon 3) inflate mRNA counts; cathepsin E-mediated shedding; apoptosis-induced survival bias in scRNA-seq data |
| LTA | LTA | Obligate heteromeric complex with LTB: the dominant form (LTα1β2) requires LTB co-expression and signals through LTBR, not TNFR1/2 | Mathematical collinearity with TNFA in ridge regression (LTA3 homotrimer binds the same TNFR1/2 receptors as TNF-α); 7 known splice variants; low/transient expression |
| HGF | HGF | Obligate mesenchymal-to-epithelial paracrine topology: HGF produced by fibroblasts/stellate cells, MET receptor on epithelial cells | Secreted as inactive pro-HGF requiring proteolytic cleavage by HGFAC/uPA (post-translational activation is rate-limiting); ECM/heparin sequestration creates stored protein pool invisible to transcriptomics |
Key insight: None of these targets have isoforms or subunits mapping to different gene IDs that would cause gene ID confusion. The poor correlations are driven by post-translational regulation (membrane shedding, proteolytic activation, decoy receptor sequestration), paracrine signaling topology (producer and responder cells are different cell types), and heteromeric complex dependence (LTA requires LTB). These represent fundamental limitations of using ligand mRNA abundance to predict downstream signaling activity — the CytoSig activity scores themselves remain valid readouts of pathway activation in the measured cells.
4.4 Cross-Atlas Consistency
4.5 Effect of Aggregation Level
Aggregation levels explained: Pseudobulk profiles are aggregated at increasingly fine cell-type resolution. At coarser levels, each pseudobulk profile averages more cells, yielding smoother expression estimates but masking cell-type-specific signals. At finer levels, each profile is more cell-type-specific but based on fewer cells.
| Atlas | Level | Description | N Cell Types |
|---|---|---|---|
| CIMA | Donor Only | Whole-sample pseudobulk per donor | 1 (all) |
| Donor × L1 | Broad lineages (B, CD4_T, CD8_T, Myeloid, NK, etc.) | 7 | |
| Donor × L2 | Intermediate (CD4_memory, CD8_naive, DC, Mono, etc.) | 28 | |
| Donor × L3 | Fine-grained (CD4_Tcm, cMono, Switched_Bm, etc.) | 39 | |
| Donor × L4 | Finest marker-annotated (CD4_Th17-like_RORC, cMono_IL1B, etc.) | 73 | |
| Inflammation | Donor Only | Whole-sample pseudobulk per donor | 1 (all) |
| Donor × L1 | Broad categories (B, DC, Mono, T_CD4/CD8 subsets, etc.) | 18 | |
| Donor × L2 | Fine-grained (Th1, Th2, Tregs, NK_adaptive, etc.) | 65 | |
| scAtlas Normal | Donor × Organ | Per-organ pseudobulk (Bladder, Blood, Breast, Lung, etc.) | 25 organs |
| Donor × Organ × CT1 | Broad cell types within each organ | 191 | |
| Donor × Organ × CT2 | Fine cell types within each organ | 356 |
4.6 Representative Scatter Plots
4.7 Biologically Important Targets Heatmap
How each correlation value is computed: For each (target, atlas) cell, we compute Spearman rank correlation between predicted cytokine activity (ridge regression z-score) and target gene expression across all donor-level pseudobulk samples. Specifically:
- Pseudobulk aggregation: For each atlas, gene expression is aggregated to the donor level (one profile per donor or donor × cell type).
- Activity inference: Ridge regression (
secactpy.ridge, λ=5×105) is applied using the signature matrix (CytoSig: 4,881 genes × 43 cytokines; SecAct: 7,919 genes × 1,249 targets) to predict activity z-scores for each pseudobulk sample. - Correlation: Spearman ρ is computed between the predicted activity z-score and the original expression of the target gene across all donor-level samples within that atlas. A positive ρ means higher predicted activity tracks with higher target gene expression.
GTEx/TCGA use donor-only pseudobulk; CIMA uses donor-only; Inflammation uses donor-only; scAtlas uses donor × organ.
4.8 Bulk RNA-seq Validation (GTEx & TCGA)
5. CytoSig vs LinCytoSig vs SecAct Comparison
5.1 Method Overview
| Property | CytoSig | LinCytoSig | SecAct |
|---|---|---|---|
| Targets | 43 cytokines | 178 (45 cell types × 1–13 cytokines) | 1,249 secreted proteins |
| Specificity | Global (cell-type agnostic) | Cell-type specific | Global |
| Source | Experimental bulk RNA-seq | CytoSig stratified by cell type (full methodology) | Spatial Moran's I |
| Best for | General cytokine activity | Cell-type-resolved analysis | Broad secretome profiling |
Six methods compared on identical matched pairs:
- CytoSig — 43 cytokines, 4,881 curated genes, cell-type agnostic (pooled from all cell types)
- LinCytoSig (orig) — cell-type-matched signatures from the CytoSig database, all 19,918 genes
- LinCytoSig (gene-filtered) — same cell-type-matched signatures, restricted to CytoSig’s 4,881 curated genes
- LinCytoSig (best-bulk) — for each cytokine, select the single best-performing cell-type signature based on GTEx+TCGA bulk RNA-seq correlation (all 19,918 genes)
- LinCytoSig (best-bulk+filt) — same best-bulk selection, restricted to CytoSig’s 4,881 genes
- SecAct — 1,249 secreted protein signatures (Moran’s I spatial method), shown for the subset of CytoSig-overlapping targets
Key findings: SecAct achieves the highest median ρ across all atlases. CytoSig outperforms the cell-type-matched LinCytoSig (orig) across all three atlases, largely because LinCytoSig signatures have fewer experiments (3–12 vs 50–300+) and more genes (19,918 vs 4,881), amplifying noise. The “best-bulk” selection strategy (selecting one representative cell-type signature per cytokine based on GTEx+TCGA bulk correlation) substantially improves performance, approaching or exceeding CytoSig. Gene filtering helps (orig < filt) consistently, confirming that restricting to CytoSig’s curated 4,881 genes reduces noise.
5.2 When Does LinCytoSig Outperform CytoSig?
LinCytoSig wins: Basophil (+0.21), NK Cell (+0.19), Dendritic Cell (+0.18)
CytoSig wins: Lymphatic Endothelial (−0.73), Adipocyte (−0.44), Osteocyte (−0.40), PBMC (−0.38)
Recommendation: Use LinCytoSig for cell-type-resolved questions and CytoSig for donor-level questions.
5.3 SecAct: Breadth Over Depth
- Highest median ρ in organ-level analyses (scAtlas normal: 0.307, cancer: 0.363)
- Highest median ρ in bulk RNA-seq (GTEx: 0.395, TCGA: 0.415)
- 97.1% positive correlation in TCGA
5.4 LinCytoSig Specificity Deep Dive
6. Key Takeaways for Scientific Discovery
6.1 What CytoAtlas Enables
- Quantitative cytokine activity per cell type per disease
- Cross-disease comparison — same 44 CytoSig signatures across 20 diseases, 35 organs, 15 cancer types
- Independent perturbation comparison — parse_10M provides 90 cytokine perturbations × 12 donors × 18 cell types for independent comparison with CytoSig predictions
- Drug-cytokine interaction — Tahoe-100M maps 95 drugs × 50 cancer cell lines
- Spatial context — SpatialCorpus-110M maps cytokine activity to spatial neighborhoods
6.2 Limitations
- Linear model: Cannot capture non-linear cytokine interactions
- Transcriptomics-only: Post-translational regulation invisible
- Signature matrix bias: Underrepresented cell types have weaker signatures
- Validation metric: Expression-activity correlation underestimates true accuracy
6.3 Future Directions
- scGPT cohort integration (~35M cells)
- cellxgene Census integration
- Drug response prediction models
- Spatial cytokine niches
- Treatment response biomarkers
7. Appendix: Technical Specifications
A. Computational Infrastructure
- GPU: NVIDIA A100 80GB (SLURM gpu partition)
- Memory: 256–512GB host RAM per node
- Pipeline: 24 Python scripts, 18 pipeline subpackages (~18.7K lines)
- API: 262 REST endpoints across 17 routers
- Frontend: 12 pages, 122 source files, 11.4K LOC
B. Statistical Methods
- Activity inference: Ridge regression (λ=5×105, z-score normalization, permutation-based significance)
- Correlation: Spearman rank correlation
- Multiple testing: Benjamini-Hochberg FDR (q < 0.05)
- Bootstrap: 100–1000 resampling iterations
- Differential: Wilcoxon rank-sum test with effect size